How to Solve the Labeled Data Problem in Industrial Automation and Robotics


Today we are witnessing a revolution in deploying robotic systems to automate repetitive tasks. We have startups trying to revolutionize everything from construction to security and agriculture using automated systems. In many cases, existing equipment is retrofitted with hardware and software for automation which of course is cheaper.

Training an automated system in robotics is a daunting task. One of the main reasons is that real world is very complex. Human brain has been trained over millennia to understand every nuance of the world. We also have the benefit of previous experience. Importantly, we have language to further augment our understanding of the world.

Automated machines, however, need to be trained explicitly. This leads to the number one problem in computer vision - lack of accurate labeled data.

Deep neural networks provide the best results in perception. However, the big challenge with traditional DNNs is that they need a lot of training data. Now, unlabeled data is relatively easy to come by. For almost any situation, we have lots of image data in the public domain. The challenge is that they are not labeled.

So how are current computer vision systems trained. With the help of humans. Most AI projects involve a dedicated team of real humans who are trained to tag and annotate images. This is a car, this one is a person and so on. However, this obviously is a very limited approach. It would take a lot of man hours to get this data. Moreover, in many cases, this requires experts such as in the case of medical imagery. This becomes cost prohibitive.

Luckily, things are changing. New developments in unsupervised learning are making in possible to do away with the need for such a large amount of labeled data. One remarkable development here is related to Generative Adversarial Networks (GANs). This technology allows us to understand images without labels better. This undrestanding can be used to generate lots of labeled data. This becomes especially useful in corner cases where it is hard to obtain data, say how to drive in the snow on a road where it rarely snows.

At reflective.ai, we work with companies to help with their data needs. We generate labeled and unlabeled data for companies to accelerate their perception projects. This could be regular images, RF signals, 3D imagery and more. Contact us at info@reflective.ai for more information.


0 views
Access Our White Paper on
Covid 19 Health & Safety Monitoring